Compute scarcity flagged again
A commentary argued that public claims about growing compute capacity actually highlight ongoing scarcity, suggesting the promised surge in accessible AI infrastructure hasn't materialized at scale. The piece uses recent OpenAI messaging as the example for why compute constraints still shape what startups can realistically build. (futurism.com)
“Compute” is the specialized chip power that trains and runs artificial intelligence models, and OpenAI is still saying it does not have enough of it. (aol.com) On April 2, OpenAI chief financial officer Sarah Friar said the company was making “very tough trades” and not pursuing some opportunities because it lacked compute capacity. Reports last week said the shortage was forcing OpenAI to turn away business in 2026. (aol.com) At the same time, OpenAI has spent months advertising much larger infrastructure plans. In September 2025, OpenAI, Oracle and SoftBank said Stargate had nearly 7 gigawatts of planned capacity and more than $400 billion in investment over three years. (openai.com) OpenAI also announced a $38 billion Amazon Web Services partnership in November 2025, with capacity targeted for deployment before the end of 2026. A month earlier, OpenAI and Advanced Micro Devices said they planned 6 gigawatts of Advanced Micro Devices graphics processing unit deployments, starting with 1 gigawatt in the second half of 2026. (aboutamazon.com) (openai.com) Those numbers describe planned buildouts, not chips already available to startups this spring. The first 1 gigawatt in the Advanced Micro Devices deal does not begin until the second half of 2026, and Amazon said its OpenAI capacity is still being deployed through the end of 2026. (openai.com) (aboutamazon.com) OpenAI has also been revising the scale of its spending story. CNBC reported on February 20 that the company told investors it was targeting about $600 billion in total compute spend by 2030 after earlier public talk of $1.4 trillion in infrastructure commitments. (cnbc.com) The bottleneck is not just software. Running large models requires graphics processing units, data centers, electricity, cooling systems and cloud contracts, and OpenAI’s own executives have tied the company’s product choices to those limits. (aol.com) (openai.com) That squeeze has shown up in project decisions. Reports on Friar’s comments said OpenAI had pulled back from efforts including the Sora video app as it redirected resources toward core products that already had strong demand. (aol.com) OpenAI has kept telling investors that more capacity will become an advantage later. Bloomberg reported on April 9 that OpenAI told investors its early push to secure compute could give it an edge over Anthropic, even as the company is still rationing opportunities now. (bloomberg.com) So the current picture is two timelines at once: multibillion-dollar infrastructure promises for 2026 and beyond, and a company in April 2026 still saying available compute is scarce enough to limit what it can build and sell. (openai.com) (aol.com)
